---
title: "AI implementation, step by step"
url: "https://mikkeltschentscher.dk/en/services/ai/ai-implementation"
language: "en"
updated: "2026-11-01"
---

# AI implementation, step by step

AI implementation is about getting AI into the way work is actually done, and the technology is rarely the hard part. I begin with one area, measure the effect and only expand if the numbers hold. After 400+ projects I know where it typically breaks.

[Get a quote](https://mikkeltschentscher.dk/en/services/ai/ai-implementation#inquiry "Let’s have a no-strings conversation")[+45 22 51 31 79](tel:+4522513179 "Call Mikkel Tschentscher at +45 22 51 31 79")

-   I usually reply the same day
-   Written estimate before work starts

![Mikkel Tschentscher](https://mikkeltschentscher.dk/images/derived/speaking.ae4c8c8ab739.png)

star reviews

[Read every review](https://mikkeltschentscher.dk/en/reviews)

-   Dedicated and capable like no one else
    
    > I have never come across a developer as dedicated and capable as Mikkel.
    
    Jonas Krogslund(translated from Danish)
    
-   Helpful and lightning fast to reply
    
    > Mikkel is seriously good, helpful, lightning fast to reply, and on top of that he is a nice person. I can warmly recommend Mikkel to others..
    
    Trine Handskemager(translated from Danish)
    
-   A fast, trouble-free process
    
    > Mikkel delivered competent, professional work setting up my website. Everything ran quickly and without a hitch.
    
    Sisse Find(translated from Danish)
    

## AI implementation

### How I approach it

The first step is finding the work that gets done many times a week, follows a fixed pattern and annoys the people doing it. That is nearly always where the first win sits, and where people will actually adopt the solution.

Next we measure how long the task takes today. Without that number, nobody can decide afterwards whether anything improved, and the discussion ends up as one gut feeling against another.

### Where AI typically gets adopted

-   Customer service: faster replies by reusing knowledge from earlier cases, with drafts that only need a touch-up.
-   Document handling: summaries and extraction from long documents, proposals and contracts.
-   Internal tools: looking things up in handbooks, procedures and product data without asking a colleague.
-   Integration with existing systems: AI features inside the programs people already use, rather than a new tab on the side.

### Guidelines and staff

Many companies either have no AI guidelines or a memo nobody has read. Both end with people using AI anyway, just without it being agreed.

I help write something short and usable: what may be entered where, what needs checking, and who to ask. Two pages people actually read beat twenty on the intranet.

### How to implement AI step by step

-   Pick one concrete task: something that happens many times a week and follows a fixed pattern. Measure how long it takes today.
-   Check the data: the data the solution needs has to be retrievable and trustworthy. If it is not, the cleanup becomes part of the plan.
-   Test on real examples: a pilot on your own cases shows what the solution can do before anyone has to rely on it day to day.
-   Agree when a person takes over: what the AI may handle on its own, and what always has to be checked.
-   Build it into daily work: the solution should sit in the systems people already work in, not in a new tab on the side.
-   Train the people using it: a short walkthrough of the solution and two pages of guidelines that people actually read.
-   Measure and adjust: compare with the baseline and fix what is not working. Make sure someone owns the solution after launch, and let the numbers decide whether to do more.

### What is AI implementation?

AI implementation is the work of going from having access to AI to using it in daily operations. It covers choosing the right first area, integrating into the systems staff already work in, training, guidelines and measuring whether the effort actually paid off. The tools exist; implementation is making them useful in your specific business.

The need exists in most companies that have bought access to ChatGPT, Copilot or similar and can see that usage is random and unevenly spread. Some employees save hours, others never touch it, and nobody knows what it adds up to. Implementation is the difference between a subscription and a gain.

An implementation partner who can build makes a practical difference. When it turns out the gain requires an integration with your line-of-business system or a small internal tool, I write it myself instead of passing the task along. I have been a developer since 2010, and that background is what the advice stands on.

Concretely, I start with conversations with the people who handle the tasks daily, find one area with volume and pattern, and put a solution into production in weeks. Then we decide from the numbers whether to do more. You own everything produced along the way, including prompts and configuration.

### AI strategy: from a list of ideas to a prioritised plan

An AI strategy does not need to be a long document. For most companies that want to implement AI, the most useful version is a prioritised list: which tasks you want to use AI for, in which order, and what you deliberately leave alone. The long version with visions and guiding principles rarely gets read, while a one-page list can be used to make decisions every month.

The list of ideas comes from conversations with the people doing the work, not from a management brainstorm. Each idea is assessed on the same few questions: How often does the task happen, and how long does it take? Does the data the solution needs exist, and can it be retrieved? What does a mistake cost? And is there someone who will own the solution afterwards? That gives an assessment of each idea that can be compared across departments.

At the top of the list there is typically a task with high volume, data that is easy to get at, and low consequences when something goes wrong. That becomes your first project, and the rest waits until the numbers from the first one are in. The plan should be revisited a couple of times a year, because both your needs and the tools change, and an idea that was too hard last year may be straightforward today.

### AI skills and change management

Since 2 February 2025, the EU AI Act has required companies that use AI at work to take steps so that their staff have a sufficient level of AI literacy. The requirement is in Article 4 and also covers entirely ordinary tools such as ChatGPT and Copilot. What counts as sufficient depends on each person's role and what the tools are used for, but the core is the same: people need to know what the tools can do, where they go wrong, and what may be entered into them.

In practice, skills are not something one course takes care of. What works is training on the tasks people actually do, one or two AI ambassadors in each team whom colleagues can ask, and management that uses the tools itself. The ambassadors are not experts, just colleagues who have got a little further and are happy to show the others how they do it. I can run the training as part of the implementation, or as an AI workshop for the whole team.

The part most often skipped is the conversation about what the time AI frees up should be used for. If nobody says it out loud, staff guess for themselves, and the guess is rarely positive. Say it early and concretely: which tasks disappear, what people will do instead, and who to go to with concerns. That is change management in its simplest form, and it matters more for whether the solution gets used than the technology itself.

### How to measure the effect of AI

The starting point is the number you measured before the solution arrived: how long the task takes and how often it happens. Afterwards you measure the same thing the same way, ideally over a few weeks, so a single busy or quiet week does not decide it. Remember to count the time staff spend checking and correcting the AI's work. It easily drops out of the calculation and can eat a large part of the gain.

Time is rarely the only thing that moves. Also keep an eye on the error rate, meaning how many cases need correcting or get sent back, on the response time to customers and, where it makes sense, on customer satisfaction. A solution that saves a little time but halves the number of errors can be worth more than one that saves a lot of time and introduces new mistakes.

Finally, measure whether the solution gets used. How many use it each week, and does usage drop after the first couple of weeks? Low usage is the earliest sign that something is wrong, typically that the solution sits in the wrong place or is not good enough yet. I gather the numbers in a short summary, so the decision to expand rests on results rather than gut feeling.

### AI implementation for companies across Denmark

Implementation is about people, so I am happy to come to you. Based in Glostrup, companies in Copenhagen and the surrounding area are easy to visit along the way, and for clients in Aarhus, Odense and Aalborg I schedule visits around kickoff and the big milestones and handle the rest over video.

The mix works well in practice: workshops and training work best in person, while follow-up, adjustment and the technical work run fine at a distance. Several of my clients I have never visited more than twice.

Typical engagements are a customer service department getting AI into their reply workflow, a finance team drowning in manual entry, or a company that has bought licences and now wants control of usage, guidelines and returns.

Two things differ from a consultancy: no junior gets sent out after the first meeting, and I have no platform or licences I need to sell. The recommendations rest solely on what benefits you.

### Why many AI projects stall

The technology is rarely the cause. The models have become good enough for most ordinary tasks, and the technical work is manageable. Projects die somewhere else.

The most common cause is that the solution sits outside the systems people work in daily. A new tab you have to remember to open does not get opened. After a fortnight, usage has dropped to the three enthusiasts who helped build it. Not out of reluctance. Nobody has room in their head for one more thing.

The second most common is that nobody owns the project after launch. The project group dissolves at go-live, reality changes with new products and new rules, and there is nobody to adjust. Quality drops slowly, people stop trusting the solution, and then it is effectively dead.

The third is expectations. If the pilot was sold internally as something that would halve a department's workload, a twenty percent improvement feels like a defeat, even though twenty percent is a great deal. I get expectations written down before the start, precisely because they otherwise grow on their own.

## What you get

Concrete deliverables, not a report of recommendations someone else has to find time for.

-   A map
    
    An overview of where in the business there is genuinely something to gain, based on talking to the people doing the work.
    
-   A measured baseline
    
    A number for how long the task takes today and how often it happens, so the effect can be assessed afterwards.
    
-   A first solution in production
    
    One area covered in four to six weeks and used by real people, not a platform meant to do everything from day one.
    
-   Guidelines
    
    Two pages on what may go where and what needs checking, short enough to actually get read.
    

## How it works

Four steps. You know what happens when, and you can stop after any of them.

1.  01 / 04
    
    ### Pick the process
    
    Together we find the piece of work that repeats often enough for automation to pay off. The more contained the task, the faster the first version reaches production.
    
2.  02 / 04
    
    ### Set a baseline
    
    We record how long the task takes today, how often it occurs, and how many errors slip through. Without those numbers, nobody can tell afterwards whether the solution was worth the effort.
    
3.  03 / 04
    
    ### Build against a test set
    
    I collect real examples from your day-to-day and test every prompt and model change against them. That keeps quality stable when the solution is tuned later on.
    
4.  04 / 04
    
    ### Production with logs and a kill switch
    
    The solution is wired into your existing systems with full logging and a switch that can turn it off. I document the setup and train someone on your side to maintain it.
    

## Brands I've worked with

-   ![MT Højgaard Danmark](https://mikkeltschentscher.dk/logos/mt-hoejgaard.svg)
    
-   ![Egmont](https://mikkeltschentscher.dk/logos/egmont.svg)
    
-   ![NNIT](https://mikkeltschentscher.dk/logos/nnit.svg)
    
-   ![Visma](https://mikkeltschentscher.dk/logos/visma.svg)
    
-   ![Lomax](https://mikkeltschentscher.dk/logos/lomax.svg)
    
-   ![Pascal](https://mikkeltschentscher.dk/logos/pascal.svg)
    
-   ![able.](https://mikkeltschentscher.dk/logos/able.svg)
    
-   ![OOONO](https://mikkeltschentscher.dk/logos/ooono.svg)
    
-   ![Novo Nordisk Fonden](https://mikkeltschentscher.dk/logos/novo-nordisk-fonden.svg)
    
-   ![Energii](https://mikkeltschentscher.dk/logos/energii.svg)
    
-   ![Maersk Tankers](https://mikkeltschentscher.dk/logos/maersk-tankers.svg)
    
-   ![Nordkysten Entreprenørfirmaet](https://mikkeltschentscher.dk/logos/nordkysten.svg)
    

## Frequently asked questions

What does AI implementation mean in practice?

Going from having access to AI to using it in daily work. It covers choosing the area, integrating into the systems people already use, training, and measuring whether it helps. The technology itself is rarely the bottleneck.

What is AI, and how does it work?

The AI tools companies use today are built on language models trained on huge amounts of text, which is why they can understand and write language, summarise, classify and extract information. They do not think; they predict the most likely answer, which is why they need the right data and a check before being used for anything that matters.

How do you create an AI strategy?

You gather ideas from the people doing the work, assess each one on volume, time spent, data and risk, and end up with a prioritised one-page list. The most important part is picking one area to start with and deliberately letting the rest wait. A good AI strategy is short enough to get used and is revisited once the numbers from the first project are in.

What does AI implementation cost?

Scope decides the price. The number of workflows, the amount of data cleanup and the need for integrations and training are the biggest items. Tell me briefly where you stand and I will reply the same day, with a concrete quote once the scope is clarified.

How big do we need to be for this to make sense?

There is no lower limit on headcount. What matters is whether a piece of work repeats often enough, and a ten-person company can easily have one.

Can we start with a pilot?

That is exactly how I recommend doing it. One area over four to six weeks, and a real decision afterwards about whether to do more. Large platform projects meant to cover everything from day one almost always end badly.

How do we measure whether AI makes a difference?

By measuring the task before the solution is adopted, and again afterwards in the same way. Time per task is the obvious number, but error rate, response time and how many people actually use the solution say at least as much. Remember to include the time spent checking the AI's work, or the gain will look bigger than it is.

What does management need to decide?

Management needs to decide which area comes first, which tools are approved and which data must never be used. It also has to appoint someone to own the solution after launch, and say openly what the time AI frees up will be used for. The rest can be delegated, but those decisions do not get made if management does not make them itself.

What about GDPR and the EU AI Act?

The guidelines we write together account for both: what may be entered where, and how output gets checked. For most ordinary uses the requirements are manageable. I am not a lawyer and will say so if you should involve one.

Do you visit companies across the whole country?

Yes, when it makes sense. Implementation involves people, so I prioritise meeting the team, especially at kickoff. Around Copenhagen that is easy, and for Aarhus, Odense or Aalborg we plan visits around the important milestones.

Do we need our data in order first?

Not all of it. But the data the first solution needs has to be retrievable and reasonably trustworthy. If it is not, the cleanup becomes part of the project and belongs in the timeline rather than arriving as a surprise.

Can we use what we already pay for?

Often yes. With Microsoft 365 or Google Workspace, AI features come with the subscription that cover a fair range of ordinary needs. I check that before proposing to build anything new.

Your next step

## Get a quote for AI implementation

Tell me about your project. I’ll get back to you with questions, options and a concrete next step.

![Mikkel Tschentscher](https://mikkeltschentscher.dk/images/derived/speaking.ae4c8c8ab739.png)

### Mikkel Tschentscher

Developer, consultant and instructor

-   16+ years of experience
-   Work directly with me
-   From advice to implementation

[+45 22 51 31 79](tel:+4522513179)[hello@mikkeltschentscher.dk](mailto:hello@mikkeltschentscher.dk)

Mikkel Tschentscher ApS · CVR 46886712

star reviews

## Related

-   [AI consultant](https://mikkeltschentscher.dk/en/services/ai/ai-consultant)
-   [AI automation](https://mikkeltschentscher.dk/en/services/ai/ai-automation)
-   [AI workshop](https://mikkeltschentscher.dk/en/services/workshops/ai-workshop)
-   [AI development](https://mikkeltschentscher.dk/en/services/ai/ai-development)

Original: https://mikkeltschentscher.dk/en/services/ai/ai-implementation
